Hybrid Mean Field Learning in Large-Scale Dynamic Robust Games
Résumé
One of the objectives in distributed interacting multi-player systems is to enable a collection of different players to achieve a desirable objective. There are two overriding challenges to achieving this objective: The first one is related to the complexity of finding optimal solution. A centralized algorithm may be prohibitively complex when there are large number of interacting players. This motivates the use of adaptive methods that enable players to self-organize into suitable, if not optimal, alternative solutions. The second challenge is limited information. Players may have limited knowledge about the status of other players, except perhaps for a small subset of neighboring players. The limitations in term of information induce robust stochastic optimization, bounded rationality and inconsistent beliefs. In this work, we investigate asymptotic pseudo-trajectories of large-scale dynamic robust games under various COmbined fully DIstributed PAyoff and Strategy Reinforcement Learning (CODIPAS-RL) under outdated noisy measurement and random updates. Extension to continuous action space is discussed.
Origine | Fichiers produits par l'(les) auteur(s) |
---|
Loading...